From Description to Causation Department of Government London School of Economics and Political Science
1 Correlation vs. Causation 2 Over-Time Changes 3 Getting Systematic about Causality
What makes something a cause? Write for 1 minute.
1 Correlation vs. Causation 2 Over-Time Changes 3 Getting Systematic about Causality
Correlation Correlation is the non-independence of two variables for a set of observations
Correlation Synonyms: correlation, covariation, relationship, association Any correlation is potentially causal X might cause Y Y might cause X X and Y might be caused by Z X and Y might cause Z There may be no causal relationship
Flashback! Two Categories of Inference: 1 Descriptive Inference What are the facts? 2 Causal Inference Why does something occur?
Correlation is Causation? The mind tends to interpret correlations and patterns as evidence of causal relationships! But this is rarely correct!
Source: Wikimedia Commons
Source: Wikimedia Commons
Source: The Economist, 8 July 2016
Source: The Economist, 8 July 2016
Source: Randal Olson
Source: Ministry of Justice, Statistics on Race and the Criminal Justice System 2010
Source: StackExchange
Source: TylerVigen.com
Naive Causal Inference Correlations are not necessarily causal Our mind thinks they are because humans are not very good at the kind of causal inference problems that social scientists care about Instead, we re good at understanding physical causality
Physical causality Action and reaction
Physical causality Action and reaction Example: Picture a ball resting on top of a hill What happens if I push the ball?
Physical causality Action and reaction Example: Picture a ball resting on top of a hill What happens if I push the ball? Features: Observable Single-case Deterministic Monocausal
1 Correlation vs. Causation 2 Over-Time Changes 3 Getting Systematic about Causality
Pre-Post Change Heuristic Our intuition about causation relies too heavily on simple comparisons of pre-post change in outcomes before and after something happens No change: no causation Increase in outcome: positive effect Decrease in outcome: negative effect
Pre-Post Change Heuristic Our intuition about causation relies too heavily on simple comparisons of pre-post change in outcomes before and after something happens No change: no causation Increase in outcome: positive effect Decrease in outcome: negative effect Why is this flawed?
Threats to Validity Campbell and Ross talk about six threats to validity (i.e., threats to causal inference) related to time-series analysis
Flaws in causal inference from pre-post comparisons 1 Maturation or trends 2 Regression to the mean 3 Selection 4 Simultaneous historical changes 5 Instrumentation changes 6 Monitoring changes behaviour
Maturation or trends Is a shift in an outcome before and after a policy change the impact of the policy or a small part of a longer time trend? Case Study: Connecticut crackdown on speeding (1955)
Regression to the mean Is a shift in an outcome before and after a policy change the impact of the policy or a function of statistical variation? Case Study: Connecticut crackdown on speeding (1955)
Selection Is a shift in an outcome before and after a policy the impact of the policy or the result of the policy being implemented when outcomes are extreme? Case Study: Connecticut crackdown on speeding (1955)
Simultaneous changes Is the shift in an outcome before and after a policy the impact of the policy or the result of a simultaneous historical shift? Case Study: US Great Depression Policy
Instrumentation changes Is the shift in an outcome before and after a policy the impact of the policy or a change in how the outcome is measured? Case Study: Age-adjusted mortality rates
Monitoring changes behaviour Is the shift in an outcome before and after a policy the impact of the policy or a change in response to measuring the outcome per se? Case Study: Educational testing
The more any quantitative social indicator is used for social decision-making, the more subject it will be to corruption pressures and the more apt it will be to distort and corrupt the social processes it is intended to monitor. Donald T. Campbell (1979)
Flaws in causal inference from pre-post comparisons 1 Maturation or trends 2 Regression to the mean 3 Selection 4 Simultaneous historical changes 5 Instrumentation changes 6 Monitoring changes behaviour
1 Correlation vs. Causation 2 Over-Time Changes 3 Getting Systematic about Causality
Directed Acyclic Graphs Causal graphs (DAGs) provide a visual representation of (possible) causal relationships
Directed Acyclic Graphs Causal graphs (DAGs) provide a visual representation of (possible) causal relationships Causality flows between variables, which are represented as nodes Variables are causally linked by arrows Causality only flows forward in time
Directed Acyclic Graphs Causal graphs (DAGs) provide a visual representation of (possible) causal relationships Causality flows between variables, which are represented as nodes Variables are causally linked by arrows Causality only flows forward in time Nodes opening a backdoor path from X Y are confounds Selection bias or Confounding
Smoking Cancer
Age Environment Smoking Cancer Parental Smoking Genetic Predisposition
Age Environment Smoking Cancer Parental Smoking Genetic Predisposition
Age Environment Smoking Cancer Parental Smoking Genetic Predisposition
Causal Inference Causal inference (typically) involves gathering data in a systematic fashion in order to assess the size and form of correlation between nodes X and Y in such a way that there are no backdoor paths between X and Y by controlling for (i.e., conditioning on, holding constant) any confounding variables, Z.